Aeon Castle (Released August 27, 2025)
https://stevesokolowski.com/songs/aeon-castle/
A detailed description of the technical workflow for this song is provided there.
The models have passed human level in the past month, both in compositional ability and audio quality, since "Ghost of You" was released. I've provided a description of how there are now independent stem-separation models that can infer the instrumentation at a higher resolution from a single condensed track, and then by adding together these instrumental stems, the output becomes far superior to the original. Using this workflow completely eliminates, for example, the biggest problem - overdriven and clipped model outputs - and reduces many other issues.
Another key technological advance is that, for the first time, I was able to use music theory to achieve a "lyrics first" workflow. The instruction following is now good enough that I was able to find "Ab dorian, clarinet, accordion, flute, EDM, 2030s," among a long list of other keywords. Instead of hoping to get something decent and then shaping the song around that, one can now - after only 4 or 6 generations - get exactly what is targeted in the keyword list, see whether that works, and then modify the keyword list.
Additionally, you can see that the album art, which took 2 minutes to generate, is on par with the best human-created art.
I specifically aimed to create a mystical feel of a long lost civilization that had used magic to advance to godlike technology, and I hope you will agree that was achieved.
1,000@WrongoPhD I suspect that you're looking for traditional pop music. This was never intended to be that type of song. Music theory was used to achieve a very specific goal here.
I will note that the reason the dorian mode is rarely used is likely for the exact purpose you specified - it does not easily lend itself to outright "hooks." That's why "Billie Jean," "Wake Me Up," and "Abracadbra" switch from dorian to minor for their choruses. "Aeon Castle" strictly adheres to the mode and never lowers the 6th.
I believe that it's extremely easy to get any of these models to output "pop" music if you just put in keywords like "pop, 2020s, c major, guitar, pachabel's canon, singer songwriter" etc and do no work on it. I'm not saying that type of music is bad, just that it's so prevalent that I don't really see any value in creating more of it.
I want to push the limits to drive past what is easy for humans to do and I understand that this song is not intended to garner radio airplay.